改进GoogLeNet模型在光纤连接器端面缺陷识别中的应用

Application of improved GoogLeNet model in visual inspection of optical fiber connector end-face defects

光纤连接器作为实现光纤之间连接的重要光无源器件,其端面表面质量的好坏会影响到光纤传输性能.经典卷积神经网络模型结构较为复杂,网络参数较多,无法满足光纤连接器端面缺陷图像识别的实时性要求.为了解决上述问题,该文设计了一种改进的GoogLeNet模型,在保障模型识别准确率的同时可提升模型的推理速度.首先,提出一种轻量化的Inception结构,减少了网络参数,保留了更多的缺陷图像的细节信息.其次,由于光纤连机器端面缺陷多为小微缺陷,对纹理特征等信息的依赖较大,通过在网络模型中引入高效通道注意力机制(ECA-Net)模块,着重提取图像中的缺陷特征.最后,对GoogLeNet模型进行改进,减少网络参数,提高模型分类性能.实验结果表明,改进的GoogLeNet模型的分类精度为95.7%,每秒传输帧数(FPS)达到了173.相较于AlexNet、VGG16、VGG19和原始的GoogLeNet模型,在保证分类精度基本一致的情况下,单张图像的推理速度分别提升了90.8%、82.9%、83.2%和81.5%.

Optical fiber connector serves as an important optical passive device to realize the connection between optical fiber,and the quality of its end face surface will affect the fiber transmission performance.The classical convolutional neural network model has a more complex structure and more network parameters,which cannot meet the real-time requirements of fiber optic connector end-face defect image recognition.Aiming to tackle the above issues,this paper designs an improved GoogLeNet model to guarantee the accuracy of model recognition and improve the inference speed of the model.Firstly,a lightweight Inception structure is proposed to reduce the network parameters and retain more detailed information of the defective images.Secondly,it is found that fiber-optic connecting machine end-face defects are mostly small micro-defects,which rely more on information such as texture features.Therefore,this paper introduces an efficient channel attention network(ECA-Net)module in the network model to focus on extracting defect features in images.Finally,the GoogLeNet model is improved to reduce the network parameters and improve the model classification performance.The experimental results show that the classification accuracy of the improved GoogLeNet model is 95.7%,and the FPS reaches 173.Compared with AlexNet,VGG16,VGG19 and the original GoogLeNet model,the inference speed of a single image is improved by 90.8%,82.9%,83.2%and 81.5%,respectively,while the classification accuracy is guaranteed to remain almost the same.